THU-BPM / ReRead

The source code of paper "Read it Twice: Towards Faithfully Interpretable Fact Verification by Revisiting Evidence" at SIGIR 2023.

Read it Twice: Towards Faithfully Interpretable Fact Verification by Revisiting Evidence

PyTorch

The code is based on PyTorch 1.13+. You can find tutorials here.

Data

Download the CHEF dataset and unzip it under the src/chef directory. You can find the download links here. The resulting directory structure should look like this:

    data/
        CHEF_train.json
        CHEF_test.json
    train.py

Usage

Run the full model on CHEF dataset with default hyperparameters

python3 train.py

This will train and evaluate our model. You may use --cuda [device] to specify the GPU device to use for training and evaluation. You can find our checkpoints at Google Drive.

Acknowledgements

We thank the authors of the original CHEF dataset and the Transformers library for their contributions.